jessicanigre-droid/malagasy-sentiment-analysis

Sentiment analysis on bank customer reviews translated into Malagasy, comparing classical ML (TF‑IDF + Logistic Regression) with transformer models (mBERT, DistilBERT) on a balanced dataset. Highlights challenges of low‑resource languages and insights into cross‑lingual transfer learning. # malagasy-sentiment-analysis Sentiment analysis on bank customer reviews translated into Malagasy, comparing classical ML (TF‑IDF + Logistic Regression) with transformer models (mBERT, DistilBERT) on a balanced dataset. Highlights challenges of low‑resource languages and insights into cross‑lingual transfer learning. # Sentiment Analysis on Bank Reviews (Malagasy Language) ## Overview This project investigates sentiment analysis for bank customer reviews translated into Malagasy, a low-resource language. It compares classical machine learning methods with transformer-based models to evaluate performance on both translated and native datasets. ## Objectives - Build a balanced dataset of 1,000 Malagasy sentences (500 positive, 500 negative). - Preprocess text to handle noise, emojis, and special characters. - Train and evaluate models: - TF-IDF + Logistic Regression (baseline) - bert-base-multilingual-cased (mBERT) - distilbert-base-multilingual-cased (DistilmBERT) - Test generalization on an external set of 200 native Malagasy sentences. ## Methodology - Preprocessing: lowercasing, URL/HTML removal, emoji conversion, normalization. - Train-validation split: 80:20, stratified by label. - External test set: HuggingFace `malagasy-sentiments-corpus`. - Models trained with AdamW optimizer and weighted loss functions. ## Results | Model | Accuracy | Precision | Recall | F1-Score | |-------------------------------|----------|-----------|--------|----------| | TF-IDF + Logistic Regression | 0.5764 | 0.6548 | 0.5609 | 0.6042 | | mBERT | 0.5514 | 0.5901 | 0.7261 | 0.6511 | | DistilmBERT | 0.4612 | 0.6056 | 0.1870 | 0.2857 | - Baseline performed well on validation but dropped on external test. - mBERT achieved higher recall, favoring positive sentiment detection. - DistilmBERT underperformed, biased toward negatives. ## Challenges - Morphological complexity of Malagasy. - Limited availabi …